You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
98 lines
3.2 KiB
98 lines
3.2 KiB
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
#pragma once
|
|
|
|
#include <vector>
|
|
|
|
#include "paddle/fluid/operators/reduce_op.h"
|
|
|
|
namespace paddle {
|
|
namespace operators {
|
|
|
|
// use for loop to speed up Eigen broadcast. 4 timer faster then broadcast
|
|
template <typename DeviceContext, typename T, typename Functor>
|
|
class ReduceSumGradKernel : public framework::OpKernel<T> {
|
|
public:
|
|
void Compute(const framework::ExecutionContext& context) const override {
|
|
auto dims = context.Attr<std::vector<int>>("dim");
|
|
if (context.GetPlace().type() == typeid(platform::CPUPlace) &&
|
|
dims.size() == 1) {
|
|
auto* input0 = context.Input<Tensor>("X");
|
|
auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
|
|
auto* output = context.Output<Tensor>(framework::GradVarName("X"));
|
|
output->mutable_data<T>(context.GetPlace());
|
|
const auto* input2_d = input2->data<T>();
|
|
auto* output_d = output->data<T>();
|
|
|
|
// handle reduce_all
|
|
if (input2->dims().size() == 1 && input2->dims()[0] == 1) {
|
|
for (int64_t i = 0; i < framework::product(input0->dims()); ++i) {
|
|
output_d[i] = input2_d[0];
|
|
}
|
|
return;
|
|
}
|
|
|
|
// handle reduce by one dimension
|
|
int reduce_dim_index = dims[0];
|
|
if (reduce_dim_index < 0) {
|
|
reduce_dim_index += input0->dims().size();
|
|
}
|
|
|
|
auto& input_dim = input0->dims();
|
|
int64_t before_dim = 1;
|
|
for (int i = 0; i < reduce_dim_index; ++i) {
|
|
before_dim *= input_dim[i];
|
|
}
|
|
int64_t reduce_dim = input_dim[reduce_dim_index];
|
|
int64_t after_dim = 1;
|
|
for (int i = reduce_dim_index + 1; i < input_dim.size(); ++i) {
|
|
after_dim *= input_dim[i];
|
|
}
|
|
for (int64_t i = 0; i < before_dim; ++i) {
|
|
for (int64_t j = 0; j < reduce_dim; ++j) {
|
|
for (int64_t k = 0; k < after_dim; ++k) {
|
|
output_d[i * reduce_dim * after_dim + j * after_dim + k] =
|
|
input2_d[i * after_dim + k];
|
|
}
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
|
|
// default use Eigen broadcast
|
|
ReduceGradKernel<DeviceContext, T, Functor> kernel;
|
|
kernel.Compute(context);
|
|
}
|
|
};
|
|
|
|
struct SumFunctor {
|
|
template <typename DeviceContext, typename X, typename Y, typename Dim>
|
|
void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) {
|
|
y->device(place) = x->sum(dim);
|
|
}
|
|
};
|
|
|
|
struct SumGradFunctor {
|
|
template <typename DeviceContext, typename X, typename Y, typename DX,
|
|
typename DY, typename Dim>
|
|
void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy,
|
|
const Dim& dim, int size) {
|
|
dx->device(place) = dy->eval().broadcast(dim);
|
|
}
|
|
};
|
|
|
|
} // namespace operators
|
|
} // namespace paddle
|